Lightweight pothole detection method based on improved YOLOv5

Aiming at the problems of large model weight volume, low detection accuracy and slow detection speed in traditional road Pothole detection methods, which are not easy to be deployed in embedded mobile devices with limited computing resources, a lightweight YOLOv5s road pothole detection model based on attention mechanism was proposed. The YOLOv5s model is taken as the basic model. Firstly, the C3 module in the backbone network is combined with the SimAM attention mechanism, which is improved to the SiMAM-C3 module, which enhances the capability of pothole feature extraction without introducing additional parameters. Then, CARAFE is used to replace the original nearest neighbor interpolation method as the upsampling method to further improve the detection accuracy of the model. Finally, SIoU loss function is used instead of CIoU loss function as bounding box regression loss function to accelerate the convergence of the model. Training and testing on datasets collected from the Internet and self-made data sets. Experimental results show that: With the same data set and the same experimental equipment, the accuracy and average accuracy of the improved YOLOv5s model reached 90.7% and 89.1%, respectively, which were 20.4% and 2.8% higher than those of the YOLOv5s model, and the parameter size was reduced by 1.1MB, showing obvious improvement effect. It can meet the accuracy and real-time requirements of pothole detection in complex environment.